Qwen3-ASR vs Whisper on Apple Silicon
Qwen3-ASR and Whisper are both strong local speech-recognition choices, but they optimize for different constraints. This comparison isolates the Qwen and Whisper rows from Soniqo's broader benchmark so accuracy, speed, memory, and implementation trade-offs are easier to evaluate.
Both engines are supported: speech-swift ships Qwen3-ASR through MLX and Whisper Large-v3 Turbo through native Core ML. WhisperKit appears below only as an external benchmark reference.
Across all rows, Qwen3-ASR 1.7B 5-bit has the lowest measured WER at 1.32%, while Qwen3-ASR 0.6B 5-bit has the lowest RTF at 0.014. In the matched 200-utterance Whisper comparison, native WhisperASR is faster: RTF 0.077 versus 0.084 and cached load 1.3 s versus 7.2 s. Accuracy is one word error apart (1.73% versus 1.71% WER), and peak RSS differs by only 18 MB (445 versus 427 MB).
Measured results
RTF is wall time divided by audio duration, so lower is faster. All rows use LibriSpeech test-clean and the same text normalizer; see the methodology caveat below for fixture sizes.
| Engine | Quant. | WER% | RTF | Peak RSS |
|---|---|---|---|---|
| Qwen3-ASR 1.7B MLX | 5-bit | 1.32 | 0.027 | 1.92 GB |
| Qwen3-ASR 1.7B MLX | 8-bit | 1.52 | 0.033 | 2.7 GB |
| Qwen3-ASR 0.6B MLX | 5-bit | 1.74 | 0.014 | 1.03 GB |
| Qwen3-ASR 0.6B MLX | 8-bit | 1.82 | 0.015 | 1.3 GB |
| WhisperKit Large-v3 Turbo | FP16 | 1.71 | 0.084 | 427 MB |
| WhisperASR native CoreML | FP16 | 1.73 | 0.077 | 445 MB |
Which one should you choose?
- Qwen3-ASR 1.7B 5-bit: choose it for the best accuracy/speed result here when a 1.92 GB peak footprint is acceptable.
- Qwen3-ASR 0.6B 5-bit: choose it for the fastest balanced Qwen option—RTF 0.014 with a 1.03 GB peak footprint.
- Native Whisper CoreML: choose it for the fastest supported Whisper path, a 1.3 s cached load, a fixed CoreML bundle, or an existing Whisper workflow.
- WhisperKit: choose it when your application already depends on WhisperKit and its model-management and decoding APIs.
Methodology and limits
The benchmark uses LibriSpeech test-clean English read speech on an Apple M5 Pro with release builds. The 8-bit Qwen and Whisper rows use the first 200 utterances (about 30 minutes). The Qwen 5-bit rows use a 100-utterance subset of the same fixture, added when those variants landed.
These numbers compare these implementations on this fixture; they are not a universal multilingual ranking. Accent, noise, long-form segmentation, language mix, decoding settings, and hardware can change the result. WER is lower-is-better, and peak RSS is measured with each engine in a separate process.